On May 25, 2026, xAI opened an early beta of Grok Build, a coding agent that runs in the terminal. The post on x.ai limits it to SuperGrok subscribers and X Premium Plus subscribers. Everyone else gets an upgrade link. Install is a single remote script, followed by a sign-in.
A coding agent, in this post, is a helper that plans a software change and then edits the repository. The surface is a command-line interface, or CLI, meaning you type in a text terminal. The loop they advertise puts a person on the plan before any file changes.
The repo I keep imagining is firmware for a small heater controller. If the temperature read fails three times, the heater output has to drop and stay down. The launch page never shows that board. It does show plan mode, a diff, and subagents that can work in separate folders. Those are the habits I would want before an agent touched the cutoff.
What actually changed?
This is the first Grok Build the post describes. There is no earlier version on the page, and no model name. I am leaving both blank.
For a hard task you start in plan mode. The agent writes the steps and waits. You may approve the plan, comment on a single step, or rewrite it. Only then does execution start. After approval, xAI says every change arrives as a clean diff. A diff is the line-by-line list of edits, the same view git already gives you. Git is the tool that stores the history of a repository.
The page's sample plan rewrites install notes for headless mode, names a -p flag, and mentions a config.toml file for models and API keys. I treat that as the screenshot's storyboard, not a manual I have checked. Headless mode is the -p switch, so a script can run an agent. The post also claims full ACP support for your own bots and orchestration apps, and it never expands ACP. I will not invent the expansion.
How does the new piece work?
xAI's install line is:
curl -fsSL https://x.ai/cli/install.sh | bash
The command downloads a script from x.ai and runs it. I read that kind of script before I trust it. Afterward you sign in with SuperGrok or X Premium Plus. No separate product key is described.
In a repository, the agent is supposed to pick up rules you already wrote. AGENTS.md is the markdown file for those rules: build commands, files that must not be reformatted, a ban on guessing a register address. Plugins, hooks, skills, and MCP servers are supposed to work as they are. A plugin is an add-on. The page shows a community one, browser-review, at version 0.8.2. A hook is a script that runs at a chosen moment, such as before a write. A skill is a pack of instructions the agent can load. MCP is the Model Context Protocol, a standard way for an assistant to call tools and read project context. A local MCP server that can open a datasheet PDF is the sort of tool the claim covers.
Bigger jobs can split across subagents, helper sessions with their own assignments, running at the same time. xAI says each can sit in its own git worktree, another working folder that shares the repository history, so two edits do not share unsaved files. /feedback, typed in the tool, is how they want bug reports during the beta.
What does this look like on a real project?
In the heater repo, AGENTS.md already says the linker script is hand-tuned. The task for plan mode is short. Add a three-strike fault on the sensor read. Leave the control gains alone. Leave the linker script alone.
If the plan comes back with a step that "simplifies" the control loop, I would comment on that step and delete it before approval. xAI says both a comment and a full rewrite are allowed. That pause is the feature. After approval I would read the diff for three things: the failure counter, the line that forces the output off, and the absence of edits in the gain table.
A register map for the README can ride on a second subagent in its own worktree, so the prose edit is not sitting in the same folder as the fault-path edit. I still merge. A worktree does not review itself. The MCP server worth having here is a reader for the sensor datasheet, so a timing figure comes from the PDF. The post says existing MCP servers work. It does not say the agent knows your chip. A hook that blocks a diff to the linker script repeats the rule already in AGENTS.md. The browser plugin in the screenshot can stay uninstalled. Headless -p waits until a person has read the fault diff and flashed a board. A script should not be the first author of a heater cutoff.
How does it compare with the previous version?
No previous Grok Build is described, so there is no version delta and no model-to-model score. The comparison that fits the page is hand editing.
By hand, you plan, you edit, and you read a diff before you commit. Grok Build moves the plan into the terminal and, on its own description, waits for approval. Git still holds the history. The new piece is the agent drafting the plan and the diff, with side jobs allowed to occupy separate worktrees.
If another terminal agent already reads your AGENTS.md, skills, hooks, and MCP servers, xAI says Grok Build will too. I have not checked that on a shared repo. The page offers compatibility, not a scoreboard.
Where does it sit next to other tools a maker already uses?
Grok Build sits beside your editor and beside git. It does not replace a debugger, a serial console, or the decision to power a board. The post never says the agent can flash a part. Flashing stays on the tools you already use.
A shop that already points Claude Code or Antigravity at the same AGENTS.md can point Grok Build at it too, if the compatibility claim holds. The suggestions will not automatically match. For a cutoff circuit I would still read the diff and put a meter on the output pin.
ACP is the advertised path for a bot or an orchestration app of your own. Until the post spells the acronym out, I would not build on it. Plan mode plus a diff is the product you can use without that guess. During the beta, comments go through /feedback.
What does it cost, and who can use it today?
No monthly price is printed. SuperGrok and X Premium Plus subscribers can install and sign in. Anyone else is sent to a SuperGrok upgrade link. The text describes no trial, no free tier, no web app, and no token rate.
What the beta lists is plan mode, diffs, AGENTS.md, plugins, hooks, skills, MCP servers, parallel subagents, git worktrees, headless -p, and ACP support. Read the install script, then sign in. That is the gate on May 25.
What is still unproven?
The post is an early-beta announcement with interface pictures. It has no benchmark, no accuracy figure, and no model name. I did not install Grok Build while writing this, so I cannot tell you whether plan mode stays in bounds on a messy firmware tree.
The docs rewrite and the latency hunt are illustrations. They do not show a linker script. AGENTS.md and a hook are how you state the rule. Whether the beta obeys it is the open question, and /feedback is the channel xAI asked for. ACP is named and not defined. Headless mode is defined, by the -p flag, and it is the mode I would enable last. A script that edits a heater controller without a fresh plan can move faster than a careful person, which is not the same as moving safely.
Until a published test says otherwise, this is a workflow for a subscription you already pay for. You still review the diff. You still flash the board.
Disclosure
Disclosure: The author is a paying subscriber to ChatGPT Plus, Claude Pro, and SuperGrok and uses all three services on a daily basis. The Makers Workbench is not affiliated with OpenAI, Anthropic, xAI, Google, or any of the other major AI companies covered in our reporting. No company receives favorable editorial treatment based on the author's personal subscriptions.
Sources and image credits
- xAI announcement (x.ai)
- Images: published by xAI with the announcement.
